> If you have four stimuli that could be categorized
> as the same 'type' of stimulus, what is the
> theoretical difference, statistically speaking,
> between specifying four different stimulus vectors of
> 15 items as four different parameters in a
> deconvolution vs. specifying one one vector of 60
> items as one parameter in the model?

Actually there are at least 3 hypotheses you could test in this scenario:

H_01: The sum of effects from the four stimulus types is 0

H_02: The effect when the four stimulus types are treated as one is 0

H_03: All four stimulus effects are 0

with the first two being the ones you have already tried. They mean differently when you reject these hypotheses: In the first one the sum of the effects is significant; the 2nd one is pretty straightforward; and there is at least one significant effect if you reject the 3rd hypothesis. You need to set up a 4-row matrix for the 3rd general linear test (glt), which is an F.

Moreover, the corresponding statistic formula and degrees of freedom are also different, which explains the following observation:

> In the results I have compared for one subject,
> the cluster regions are similar, but the former case
> yields clusters of voxels significant to 1e-9,
> whereas the second only around 1e-4.

It is quite obvious that you would easily get clusters with the 1st test since it does not necessarily require each of the stimuli being active.

Regarding "which approach would be advisable", it really depends on what kind of questions you have in mind. As demonstrated by their null hypotheses, each of the above 3 tests answers a slightly different question: You get what you asked for.

Gang